Re: [PATCH v3 04/36] mtd: st_spi_fsm: Supply framework for devicerequests

From: Brian Norris
Date: Tue Dec 10 2013 - 15:19:53 EST


On Fri, Nov 29, 2013 at 12:18:53PM +0000, Lee Jones wrote:
> --- a/drivers/mtd/devices/st_spi_fsm.c
> +++ b/drivers/mtd/devices/st_spi_fsm.c
> @@ -46,6 +51,36 @@ static void stfsm_clear_fifo(struct stfsm *fsm)
> }
> }
>
> +static inline void stfsm_load_seq(struct stfsm *fsm,
> + const struct stfsm_seq *seq)
> +{
> + void __iomem *dst = fsm->base + SPI_FAST_SEQ_TRANSFER_SIZE;
> + const uint32_t *src = (const uint32_t *)seq;
> + int words = STFSM_SEQ_SIZE / sizeof(uint32_t);

I think this is clearer as:

int words = sizeof(*seq) / sizeof(*src);

> +
> + BUG_ON(!stfsm_is_idle(fsm));
> +
> + while (words--) {
> + writel(*src, dst);
> + src++;
> + dst += 4;
> + }
> +}
> +
> +static void stfsm_wait_seq(struct stfsm *fsm)
> +{
> + unsigned long timeo = jiffies + HZ;
> +
> + while (time_before(jiffies, timeo)) {
> + if (stfsm_is_idle(fsm))
> + return;
> +
> + cond_resched();
> + }
> +
> + dev_err(fsm->dev, "timeout on sequence completion\n");

I believe the timeout logic is incorrect. What if we wait a "long time"
to call stfsm_wait_seq() (due to scheduling, or otherwise)? Then the
while loop might not even run once (time_before(x, y) is false). Or what
if cond_resched() waits for a long time...

So you need an extra check of stfsm_is_idle() after the while loop,
before you declare a timeout.

> +}
> +
> static int stfsm_set_mode(struct stfsm *fsm, uint32_t mode)
> {
> int ret, timeout = 10;
> diff --git a/drivers/mtd/devices/st_spi_fsm.h b/drivers/mtd/devices/st_spi_fsm.h
> index 4e92e58..6164142 100644
> --- a/drivers/mtd/devices/st_spi_fsm.h
> +++ b/drivers/mtd/devices/st_spi_fsm.h
> @@ -199,4 +199,18 @@ struct stfsm {
> uint32_t fifo_dir_delay;
> };
>
> +struct stfsm_seq {
> + uint32_t data_size;
> + uint32_t addr1;
> + uint32_t addr2;
> + uint32_t addr_cfg;
> + uint32_t seq_opc[5];
> + uint32_t mode;
> + uint32_t dummy;
> + uint32_t status;
> + uint8_t seq[16];
> + uint32_t seq_cfg;
> +} __attribute__((__packed__, aligned(4)));

I think checkpatch.pl prefers these attributes use the kernel macros:

struct stfsm_seq {
...
} __packed __aligned(4);

> +#define STFSM_SEQ_SIZE sizeof(struct stfsm_seq)

If you agree with my earlier comment, you won't need this macro.

> +
> #endif /* ST_SPI_FSM_H */

Brian
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/